MIT researchers built an algorithm that makes generative AI models obey strict rules. The method, called HardFlow, enforces requirements such as safety limits or physical laws on the final output only. Earlier approaches forced each intermediate step to satisfy the rules, which held back solution quality. HardFlow treats the problem as one of steering the generation path, borrowing tools from control theory. It works at deployment time on already-trained models, so no retraining is needed. Tests covered robotic manipulation, maze navigation and text-guided image editing. The team reports full constraint satisfaction and better solution quality than existing methods. Computation time was similar to or lower than competing approaches. The researchers next want to extend it to cases where the model itself can be updated. Navid Azizan, Zeyang Li and Kaveh Alim published the work in IEEE Transactions on Pattern Analysis and Machine Intelligence.